基于PReLUs-Softplus非线性激励函数的卷积神经网络

摘要

Aiming at the problem that the expression ability and recognition effect of convolutional neural network ( CNN) are affected by the excitation function of convolutional layer, a new nonlinear excitation function PReLUs-Softplus was proposed and applied to the convolutional layer in neural network. The contrast experiments on the image recognition of both new neural network and neural network with the traditional excitation function were performed in MNIST and CIFAR-10 standard database. The results show that compared with the neural network with the traditional excitation function, the convolutional neural network with PReLUs-Softplus excitation function has faster convergence rate in the calculation of image recognition under different pooling methods, and can effectively reduce the recognition error rate.%针对卷积神经网络表达能力和识别效果受卷积层激励函数影响的问题,提出了一种新型非线性激励函数PReLUs-Softplus,并将其应用于神经网络卷积层.对新型神经网络和采用传统激励函数的神经网络在MNIST和CIFAR-10标准数据库上进行了图像识别对比实验,结果表明,相比于采用传统激励函数的神经网络,使用PReLUs-Softplus激励函数的卷积神经网络在不同的池化方式下图像识别计算收敛速度更快,显著降低了识别的错误率.